Re: [PATCH v2 3/5] perf report: Sort by sampled cycles percent per block for stdio

[Jin, Yao]

On 10/15/2019 4:41 PM, Jiri Olsa wrote:
> On Tue, Oct 15, 2019 at 01:33:48PM +0800, Jin Yao wrote:
>
> SNIP
>
> > +enum {
> > +	PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_COV,
> > +	PERF_HPP_REPORT__BLOCK_LBR_CYCLES,
> > +	PERF_HPP_REPORT__BLOCK_CYCLES_PCT,
> > +	PERF_HPP_REPORT__BLOCK_AVG_CYCLES,
> > +	PERF_HPP_REPORT__BLOCK_RANGE,
> > +	PERF_HPP_REPORT__BLOCK_DSO,
> > +	PERF_HPP_REPORT__BLOCK_MAX_INDEX
> > +};
> > +
> > +static struct block_fmt block_fmts[PERF_HPP_REPORT__BLOCK_MAX_INDEX];
> > +
> > +static struct block_header_column{
> > +	const char *name;
> > +	int width;
> > +} block_columns[PERF_HPP_REPORT__BLOCK_MAX_INDEX] = {
> > +	[PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_COV] = {
> > +		.name = "Sampled Cycles%",
> > +		.width = 15,
> > +	},
> > +	[PERF_HPP_REPORT__BLOCK_LBR_CYCLES] = {
> > +		.name = "Sampled Cycles",
> > +		.width = 14,
> > +	},
> > +	[PERF_HPP_REPORT__BLOCK_CYCLES_PCT] = {
> > +		.name = "Avg Cycles%",
> > +		.width = 11,
> > +	},
> > +	[PERF_HPP_REPORT__BLOCK_AVG_CYCLES] = {
> > +		.name = "Avg Cycles",
> > +		.width = 10,
> > +	},
> > +	[PERF_HPP_REPORT__BLOCK_RANGE] = {
> > +		.name = "[Program Block Range]",
> > +		.width = 70,
> > +	},
> > +	[PERF_HPP_REPORT__BLOCK_DSO] = {
> > +		.name = "Shared Object",
> > +		.width = 20,
> > +	}
> >   };
>
> so we already have support for multiple columns,
> why don't you add those as 'struct sort_entry' objects?

For 'struct sort_entry' objects, do you mean I should reuse the 
"sort_dso" which has been implemented yet in util/sort.c?

For other columns, it looks we can't reuse the existing sort_entry objects.

> SNIP
>
> > +{
> > +	struct block_hist *bh = &rep->block_hist;
> > +
> > +	get_block_hists(hists, bh, rep);
> > +	symbol_conf.report_individual_block = true;
> > +	hists__fprintf(&bh->block_hists, true, 0, 0, 0,
> > +		       stdout, true);
> > +	hists__delete_entries(&bh->block_hists);
> > +	return 0;
> > +}
> > +
> >   static int perf_evlist__tty_browse_hists(struct evlist *evlist,
> >   					 struct report *rep,
> >   					 const char *help)
> > @@ -500,6 +900,12 @@ static int perf_evlist__tty_browse_hists(struct evlist *evlist,
> >   			continue;
> >   
> >   		hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
> > +
> > +		if (rep->total_cycles) {
> > +			hists__fprintf_all_blocks(hists, rep);
>
> so this call kicks all the block info setup/count/print, right?

Yes, all in this call.

> I thingk it shouldn't be in the output code, but in the code before..
> from what I see you could count block_info counts during the sample
> processing, no?

In sample processing, we just get all symbols and account the cycles per 
symbol. We need to create/count the block_info at some points after the 
sample processing.

Maybe it's not very good to put block info setup/count/print in a call, 
but it's really not easy to process the block_info during the sample 
processing.

Thanks
Jin Yao

> jirka